A Google TechTalk, 2026-02-18, presented by Andy Dong
ABSTRACT: DP-SGD remains the standard approach for private model training. A variety of techniques have been developed to improve its privacy–utility tradeoff, including privacy amplification through data subsampling, leveraging structured randomness in the training process, and correlated-noise mechanisms such as DP matrix factorization (DP-MF). While DP-MF can improve utility, its interaction with subsampling-based amplification is less explored than in the classical DP-SGD setting.
In this talk, I present b-min-sep subsampling, a simple batching scheme for DP-MF. The key idea is to impose only a minimal participation constraint that preserves the structural properties required by correlated-noise mechanisms, while retaining substantial flexibility in sampling. The resulting scheme improves over cyclic Poisson subsampling and is a generalization of balls-in-bins subsampling.
I will give a high-level overview of the privacy analysis based on Monte Carlo accounting and dynamic programming, and present empirical results demonstrating improved privacy–utility tradeoffs in both example-level and multi-attribution user-level settings. More broadly, this work fits into a line of research on leveraging randomness in training for privacy amplification.